acceptodds
Under review as a conference paper at ICLR 2027

Survival World: Exploring Alternative Patient Futures Through Trajectory Reasoning for Survival Prediction

Abstract

Longitudinal survival prediction requires reasoning about how a patient's conditionmay evolve beyond the observed history. We introduce SURVIVAL WORLD (SW),an empirical world model that makes these possible futures explicit and traceable.The model organizes observed training transitions into an executable graph and usesa language model to compile feature semantics into a bounded retrieval program. Ateach rollout state, the program adapts retrieval to the current measurements, trends,and observation context, then advances to an actual training history. A sharedsurvival head evaluates the resulting trajectories, and averaging their survivalprobabilities integrates alternative futures into a unified prediction. Programsexecute locally, with reusable computations and no patient records sent to thelanguage model. In an exploratory evaluation on PBCseq and a MIMIC-IV cohortat the second admission, with adaptive selection using test scores, the methodachieves the highest selected concordance among seven baseline families and threeworld model controls. Component ablations identify the roles of empirical retrieval,dynamic weighting, and path aggregation. Beyond prediction, exploring alternativenode sequences reveals interpretable state profiles associated with different modeledsurvival trajectories, enabling traceable hypothesis generation. Survival World connects semantic program design with empirical patient dynamics to support both survival prediction and exploration of possible futures.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Survival Prediction,World Models,Patient Trajectory Reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.60 over 5 reviews, above 10% of the others.
Reviewer_x9YMOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents Survival World (SW), an empirical framework for survival prediction. It attempts to construct multiple alternative trajectories by retrieving and composing observed patient transitions. Path-wise cumulative-hazard contributions are then aggregated into head-specific survival curves, which are in turn combined into the final prediction. Major comments:

  1. The details of the search program, including the survival-head combinations, aggregation choices, and the complete search protocol, etc., haven’t been fully explained.
  2. It is unclear where the formula in the last line of Figure 1 comes from (how it relates to Eqs. (6) and (7)), how the feature weights (w) are computed, and how the transition graph is created.
  3. There are many confusing statements used, like “empirical futures” “state action” in Figure 1, “training-supervision” “training-fitted” ,…
  4. When no historical state provides a sufficiently close match, how does SW come up with the candidates? How reliable is the trajectory it generated then? 5.The learning algorithm is missing.
Critical strengths
  1. The proposed SW attempts to preserve the data provenance of each transition, which should make it easier to trace the predictions.
  2. Baseline comparisons and ablation studies show that SW might be more accurate on some predictive metrics.
Critical weaknesses
  1. The paper is hard to understand, mostly because it's poorly presented and a lot of details are missing.
  2. The exact meanings of key terms like “world model” and “alternative patient futures” haven't been discussed.
  3. Formal definitions for the key concepts, such as Future/Observable states, Disease trajectories, and Retrieval rules, etc., are missing.
  4. Lack of experimental validation for the effectiveness of the resulting future states and disease trajectories.
Questions for authors

Provide a well-defined learning algorithm.

Reviewer_x11fOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Survival World (SW), an empirical world model for longitudinal survival prediction that makes possible patient futures explicit and traceable. Observed training transitions are organised into an executable graph, and a language model compiles feature semantics into a bounded retrieval program. At each rollout state, this program adapts retrieval to the current measurements, trends, and observation context, and then advances to an actual training history. A shared survival head scores the resulting trajectories, and their survival probabilities are averaged into a single prediction. Programs execute locally, so no patient records are sent to the language model.

The method is evaluated on PBCseq and a MIMIC-IV cohort at the second admission against seven baseline families and three world-model controls, where SW reports the highest selected concordance. The validation is complemented by component ablations (empirical retrieval, dynamic weighting, path aggregation), calibration analysis, per-seed results for SW, and a qualitative exploration of alternative node sequences as interpretable state profiles associated with different survival trajectories.

Critical strengths
  1. Traceability with controlled leakage. The proofs of the construction properties are correct, and every rollout advances to an actual training history, so each predicted future remains traceable to observed patients. The language model is restricted to feature semantics and is never exposed to patient outcomes, which correctly isolates the endpoints. This is consequential because it makes the interpretability claim credible and removes an obvious source of leakage in an LLM-in-the-loop survival model.

  2. Competitive accuracy while remaining interpretable. SW matches or exceeds established survival baselines in concordance while exposing the alternative futures that support each prediction. The before and after calibration results further confirm that the calibration step does not introduce leakage. This supports acceptance because interpretability is obtained without a loss in predictive performance.

Additional points: reporting per-seed metric scores for SW is a welcome additional robustness check.

Critical weaknesses
  1. The practical value of interpretability is not established. Interpretability is important for ML models in general, but the paper provides no metric, study, or supporting literature showing why it is needed in this particular clinical setting or how it translates into usefulness. This is consequential because the concordance gains of SW over the strongest baselines lie in the range of 10−310^{-3} to 10−410^{-4}, and full SW does not improve significantly over the variants that remove retrieval or trajectory construction. The contribution therefore rests largely on the interpretability argument, which is currently unsupported.

  2. Variance and statistical rigour of the comparison. The ML/DL baselines are reported without seed averages or standard deviations, and the paper does not state whether they were run from existing checkpoints or retrained. For SW itself, the original patient splits and search seed are retained, so the survival heads and the LLM retrieval program come from a single training run, leaving variance under different sample selections unassessed. Given the very small margins, these factors could change the overall ranking.

Additional points:

  • The paper argues that DeepHit lacks a longitudinal component but does not include Dynamic-DeepHit, its longitudinal successor by the same authors, which is a strong and directly relevant baseline.
  • Comparisons with graph-based RAG methods and with statistical approaches designed specifically for longitudinal-to-survival prediction are also missing.
  • The computational cost of the LLM backbone (search rounds, tokens, wall-clock time during training) is not reported.
Questions for authors

Questions that could affect my assessment:

  1. Can the authors provide statistical significance tests or confidence intervals for the C-index improvements over the baselines, computed across multiple seeds?
  2. Could the authors add at least one comparison against Dynamic-DeepHit, a graph-based RAG method, or a statistical longitudinal-to-survival model?
  3. How sensitive is the model to the prognostic-coordinate weight (4 and 0) and the semantic fraction (0.5 and 1)?
  4. Could the authors add an ablation comparing the LLM-authored program against a hand-designed program of matched complexity, evaluated on the final two-head architecture across the same five seeds used for the headline results? This would clarify the actual contribution of the LLM.

Suggestions (nice to have):

  1. Related to the computational cost raised among the weaknesses, could search rounds, token usage, and wall-clock time be reported for the policies in Appendix C?
  2. Is there any preliminary evidence that the trajectory explorations in Figures 2 and 3 could change clinical decisions or judgements compared with a static prediction?
Reviewer_gPxkOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper aims to build framework that builds alternative patient trajectories by connecting observed transitions from training patients, using an LLM-generated program to guide retrieval. These trajectories are evaluated using shared survival models and combined into survival predictions, while keeping track of the original patient records supporting each transition. An LLM writes the rules that decide which transitions get retrieved, and since every step points back to a real training record, the authors argue you can trace a prediction back to the patients it came from.

Critical strengths
  • The idea of building possible patient futures from observed transitions, rather than generating completely synthetic data, is interesting.
  • Being able to trace each transition back to a real patient record is a useful feature of the approach.
  • The authors evaluate the method on two datasets, PBCseq and MIMIC-IV, and compare it with several established survival prediction methods.
  • I appreciate that the authors included several ablation experiments to better understand the different components of the method, and besides, were honest with the limitations.
Critical weaknesses
  • I found the paper quite difficult to follow. The method is already complex, but the way it is presented makes it harder to understand than necessary. There are many technical details introduced before the overall workflow becomes clear. By the end of the results section, I was still struggling to understand which experiments actually support the main claims.
  • Figure 1 needs substantial improvement and a big amount of reworking from my point of view. Some arrows point in confusing directions, text overlaps with lines, and several connections are difficult to follow, and icons are somehow misaligned or weirdly generated. The figure also tries to explain too many things at once. Since this is the main figure introducing the method, I think it needs to be redesigned.
  • The experimental section is particularly confusing. The authors switch between different model configurations, numbers of trajectories, pooling methods, and training setups. Although these differences are mentioned, it is difficult to keep track of what is being compared and why. There are many experiments, but the main findings get lost in all the details.
  • I am not convinced that the additional complexity is justified by the improvements. On MIMIC-IV, SW improves the C-index over the anchor model by only .00039. The improvement over TabM-Cox, which is also used as the survival backbone of SW, is even smaller (0.00023). It is difficult to see how meaningful these differences are.
  • Some of the ablation results raise questions about the need for the proposed components. In Table 4, removing semantic distance actually improves the PBCseq C-index. For instance, in Table 5, the prefix-trained anchor outperforms SW on MIMIC-IV across all three metrics, and the calibration-based selection assigns zero weight to the trajectory component. To me, this is a particularly important result, as the method itself ends up discarding the trajectory contribution in that setting.
  • From the results I cannot inferred how much the LLM-generated retrieval program contributes for example. Three of the four generated branches for PBCseq cannot even activate because the required information is unavailable. This makes me wonder whether the LLM is really learning useful retrieval rules or whether a much simpler approach would perform similarly.
  • lthough each transition comes from a real patient record, the complete trajectory may combine observations from several patients. This does not necessarily mean that the resulting trajectory represents a realistic disease progression. Here comes into play if this is clinically meaningful.
  • If I got it correctly, the PBCseq test set contains only 63 patients and 28 events, yet the results are reported to six decimal places without confidence intervals in the main comparison. Also, the authors mention that SW uses a substantially larger search budget than the baselines, which makes it difficult to judge how fair the comparison is.
Questions for authors
  1. Could the authors explain the method more simply, perhaps using one concrete patient example? If I understand correctly, the framework retrieves observed transitions, connects them into possible future trajectories, and then uses survival models to estimate risk along these paths. I think this could be explained much more clearly before introducing all the equations.
  2. Why are different configurations used across the main comparisons and ablations? I understand that the experiments have different purposes, but the current presentation makes it difficult to understand what conclusions can be drawn from each one.
  3. What is the actual benefit of using an LLM-generated retrieval program? In Table 4, removing semantic distance slightly improves performance, and Appendix C shows that most of the generated branches for PBCseq cannot activate. Would simpler, manually defined retrieval rules achieve similar results?
  4. How meaningful are the improvements on MIMIC-IV? The differences are extremely small, including when compared with the TabM-Cox backbone. Have the authors tested whether these improvements hold across different training splits, rather than only different simulation seeds? --> Statistically significant? what are the drawbacks?
  5. Regarding the results from Table 5, the prefix-trained anchor performs better than SW on MIMIC-IV across all three metrics, and the selected model gives zero weight to the trajectory component. Doesn't this suggest that trajectory generation is not actually helping in this setting? How do the authors interpret this result in relation to their main claims?
  6. Have the authors checked whether the generated trajectories make sense clinically? Combining real transitions from different patients could still produce unrealistic sequences. It would be useful to know whether these trajectories were evaluated beyond their effect on survival prediction.
  7. Could the authors clarify which data were used to provide feedback during the LLM program search? Also, given that SW uses a larger search budget than the baselines, how was this taken into account to ensure a fair comparison?
  8. Given the small PBCseq test set, could the authors report confidence intervals for the main comparisons? I am not sure reporting C-index values to six decimal places is particularly informative here.
  9. Have the authors considered comparing the generated trajectories with actual future observations from held-out patients? I think this would help show whether the method is really capturing patient progression, rather than simply connecting observed transitions in ways that produce different survival predictions.
Reviewer_aApzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Survival World (SW), a framework for survival prediction that explores possible future patient trajectories using observed transitions from historical patient records. Instead of learning a separate generative dynamics model, SW retrieves transitions from a training-patient graph and combines them into multiple possible future paths. An LLM is used to design retrieval programs based on feature descriptions and aggregate evaluation feedback, while shared survival predictors estimate survival probabilities along the generated trajectories. The main idea is to make survival predictions more interpretable by linking possible future developments to observed patient data. The authors evaluate the framework on PBCseq and MIMIC-IV, comparing it with ten survival prediction baselines. The reported results show some improvements in C-index, while the ablation studies examine retrieval strategies, trajectory aggregation, training supervision, and calibration. The paper also includes examples illustrating how different retrieved trajectories correspond to different survival predictions.

Critical strengths
  1. I like the motivation behind moving beyond a single survival risk score and exploring possible patient futures. Traditional survival models usually provide limited information about how a patient's condition might evolve. The attempt to connect survival prediction with explicit future trajectories is interesting, particularly for applications where understanding the prediction matters.

  2. One appealing aspect of SW is that every simulated transition is linked to an observed transition in the training data. This makes the generated scenarios easier to inspect than trajectories produced entirely by a black-box generative model. I see this as a practical advantage, even though the complete trajectories are still model-constructed.

  3. Rather than training another neural network to predict future patient states, the authors build trajectories through empirical retrieval. This is a sensible design choice, especially when longitudinal medical data are limited. The use of cached computations and reusable survival heads also makes the overall pipeline reasonably structured.

Critical weaknesses
  1. I am not fully convinced by the methodological novelty. The overall framework mainly combines nearest-neighbor retrieval, LLM-generated retrieval rules, empirical trajectory sampling, and existing survival predictors. Most of these components have already been explored in related contexts. While their combination is reasonable, I struggled to identify a sufficiently distinct learning or reasoning mechanism. In particular, calling the approach a world model seems somewhat ambitious when the underlying dynamics are constructed by stitching together observed transitions rather than learning a generalizable transition model.

  2. In Table 2, SW improves the best baseline C-index from 0.8850 to 0.8925 on PBCseq, but the improvement on MIMIC-IV is only from 0.7334 to 0.7336. Moreover, SW does not consistently outperform existing methods on Brier score or IBS. Considering the additional complexity introduced by the retrieval program and trajectory simulation, I would have expected clearer evidence that these components provide a meaningful predictive advantage.

  3. Although each retrieved edge comes from an actual patient record, a complete trajectory can combine transitions from multiple unrelated patients. There is no guarantee that these combinations represent realistic disease progression. The scenario examples are visually interesting, but they mainly demonstrate differences in model-generated predictions. I would find the interpretability claim more convincing if the authors evaluated temporal consistency, clinical plausibility, or agreement with subsequently observed patient trajectories.

  4. Results in Table 5 are especially interesting here. On MIMIC-IV, the prefix-trained anchor achieves a C-index of 0.7376, slightly higher than the corresponding SW model at 0.7374, and also performs better on the other reported metrics. This suggests that some improvements may come from better survival-head training rather than trajectory construction. Without a clearer separation between these contributions, the central claim about the value of alternative future reasoning feels somewhat under-supported.

  5. PBCseq contains only 63 test patients and 28 events, making the reported improvements potentially sensitive to individual cases. The five simulation seeds mainly assess sampling variability on the same patients rather than uncertainty across patient populations. The authors also acknowledge that the program-search budget is substantially larger than baseline tuning budgets. Finally, I could not clearly determine how the aggregate feedback used during LLM program search is separated from final test evaluation. More explicit documentation and patient-level uncertainty analysis would strengthen the results considerably.

Questions for authors
  1. How much does the LLM-generated retrieval program actually contribute? Table 4 shows that removing semantic distance slightly improves the PBCseq C-index. Could the authors explain this result and clarify whether the gains come primarily from the program's feature weights, adaptive neighborhood selection, or other retrieval parameters?

  2. Why does trajectory reasoning provide such a limited improvement on MIMIC-IV? The full model performs only marginally better than geometric retrieval and the matched anchor, while Table 5 shows that prefix supervision can benefit the anchor more than the trajectory-based predictor. Is there evidence that SW offers consistent gains for particular patient subgroups or prediction horizons?

  3. How is the independence of the test set maintained during program search? The LLM receives aggregate evaluation feedback, and the manuscript mentions substantially different search budgets. Could the authors specify which data split provides this feedback, how many candidate programs are evaluated, and whether the final configurations were selected before accessing TEST results?

Reviewer_3AENOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes Survival World (SW), a retrieval-based survival prediction framework that assembles observed patient transitions into alternative future trajectories. An LLM designs a retrieval program to select transitions, and hazard contributions from fitted TabM-Cox models are aggregated along and across trajectories to produce a survival forecast. The method is evaluated on PBCseq and a second-admission MIMIC-IV cohort through comparisons with survival baselines and ablation studies

Critical strengths

The framework links alternative survival forecasts to the retrieved patient trajectories used to construct them. This allows users to inspect the historical observations associated with different predictions.

Critical weaknesses
  1. Unclear contribution beyond conditional resampling.

The trajectory generator closely resembles nearest-neighbor and local bootstrap methods, which recursively sample observed successors of states similar to the current state (Lall and Sharma, 1996; Paparoditis and Politis, 2002). The potential contribution lies in the programmable retrieval policy and its integration with survival prediction, but their incremental value is not clearly established. Most of the PBCseq concordance gain over the matched anchor is already obtained by the geometric-rollout control in Table 3. Furthermore, in Table 5’s separate MIMIC-IV supervision experiment, the prefix-trained anchor has better reported values than its SW counterpart on all three metrics. The paper should clarify its distinction from existing conditional resampling methods and provide stronger evidence for the benefit of the additional complexity.

  1. Gap between the world-model framing and the interpretation of the rollouts.

SW is motivated as modeling the evolution of a patient’s condition. However, each rollout step adopts another patient’s entire observed history prefix, rather than extending a history consistent with the query patient’s observations. Although this defines an empirical transition mechanism, the paper does not clearly explain how the resulting sequence of donor-history substitutions should be interpreted as the evolving state of the same patient. This connection is important for understanding what the world-model framing and the claimed alternative patient futures mean beyond a sequence of retrieved examples used for survival prediction.

Reference

Lall, U., & Sharma, A. (1996). A nearest neighbor bootstrap for resampling hydrologic time series. Water Resources Research, 32(3), 679–693.

Paparoditis, E., & Politis, D. N. (2002). The local bootstrap for Markov processes. Journal of Statistical Planning and Inference, 108(1–2), 301–328.

Questions for authors

Could you revise Section 3 to make the method and notation easier to follow? Section 3 provides insufficient explanation of the key quantities. Some notations are reused for different quantities, making the presentation ambiguous.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.